Fix GenAI’s Greatest Vulnerabilities With Playful Design

Sam Liberty
6 min readJul 28, 2024

--

Whether you’ve articulated it to yourself or not, part of you realizesthat all GenAI products have the same three major vulnerabilities:

  1. Lack of Resilience
  2. Frustrating User Experience
  3. Issues of Trust

Most of the negative press you’ve seen around GenAI likely hinges around one or more of these problems.

People exploit GenAI’s lack of resilience by “hacking” it to do and say ridiculous things through creative prompting.

GenAI refuses (or fails) to answer simple questions because of its limitations.

Large language models are largely black boxes that seem like magic to the typical user, and therefore it is hard to trust what they produce, especially when they have been known to hallucinate and steal.

The key to solving all three of these vulnerabilities, and in fact turn them into strengths, is shockingly simple.

It’s play.

Play = Resilience

A truism in playful design circles is that whatever you make, people will try to break. That is, they’ll play with it. Humans are curious, and when they encounter something new, they prod at it and contort it into new configurations. We’re like raptors checking the fences for security flaws.

This is why Siri became a cultural phenomenon while Google’s Assistant was a largely forgettable tool that’s at its best when it fades into the background. The designers at Apple knew that people would want to play with their new AI assistant, so they programmed it to play back with us. In short, they gave Siri a personality, and made sure to make that personality playful. Siri uses humor and her developers embedded fun Easter Eggs into her design that creative users can find. This presentation makes unexpected results a feature, not a bug.

A product that expects to be played with is prepared for the unexpected. Play makes us more adaptable, more resilient. Just as Brian Sutton-Smith claims that the evolutionary purpose of play is to boost brain elasticity and help us adapt to new threats and opportunities as we age, AI agents need to be playful, so that when humans inevitably fuck with them, they can fight fire with fire.

Despite the negative press, LLMs are actually very robust and resilient. The problem is, their personalities are the equivalent of warm vanilla soy milk.

I asked about ChatGPT’s tone of voice…

Play gives your product a powerful set of tools it can use to counter-punch and adapt to all the bizarre an unpredictable jokes we humans will play on it.

Play = Delight

LLMs express many frustrating little quirks. They hallucinate. They over-apologize and sycophantically praise their prompters. They write in abominable tones and produce language unsuitable for almost any purpose.

When ChatGPT was first launched, part of what drew people to it was the delight of using it. Watching it create whimsical, unexpected, and fun text was truly novel. Every time you prompted it, something truly amazing might come out the other side.

Today, the novelty of spontaneously generated text has pretty much worn off. Instead, users are frustrated by the sameness and blandness in the responses delivered by these models.

I asked GPT to show me something truly shocking and it gave me an anodine science fact

What if instead, GenAI products were tuned to give us something different? What if instead of being bored by what they produced we were shocked, surprised, and blown away?

Now you might be thinking “It’s text on a gray background, how different can it be? This is another great point, so thank you for raising it.

Why do the vast majority of GenAI products all look basically the same?

They’re so focused on the differences between their models that they totally ignore the differences in their User Interface designs. That and they rightfully assume that when people ask a question, they want a straightforward, useful answer.

So if you don’t want to mess with the text itself much, what else can you do with your product? Can you generate sounds? Images? Animations? More playfully designed UI elements?

The possibilities are limitless, especially with the power of GenAI at your fingertips. So go big and differentiate your product with delight.

Play = Trust

When you play a game with somebody, you have no choice but to rapidly build trust with them, even if they are a stranger and even if they are your opponent.

In games, we trust that others will follow the rules. We also trust that our teammates will follow through to the best of their abilities, and play with us in good faith.

Trust requires us to make a commitment and then follow through on that commitment. It also requires the demonstration of competence.

I asked ChatGPT what it might say if it wanted my trust and respect, and it offered this totally normal human paragraph.

Playing a game gives us many, many opportunities to do these things in a short period of time. This is why a playful product is paradoxically more trustworthy than a “serious” one.

Evolutionary scientists even theorize that smiling and laughter are ways to signal trustworthiness and form bonds. Proof in point, the best way of getting someone to trust you is to show them a genuine smile.

GenAI might not be great at smiling, but it can certainly play if properly designed.

But trust is more than a playful interface. The inner workings of LLMs are mysterious to most. But people trust what they themselves have made, like a climber who trusts a knot they tied themself.

I asked ChatGPT to name any individuals who helped design educational features of the product, and it could not name any.

This is why the design of AI products should also be robust and collaborative, and involve a working group made up of the actual customers. If an AI agent is meant as a teaching tool, it should be designed in cooperation with teachers… and not just one or two.

In fact, if I were asked to design an AI as a teaching tool for a k12 school, the first thing I would do is make sure I had the buy-in every single teacher and administrator by involving them in the design process in whatever capacity they’re willing to partake.

Personality goes a long way.

Playful Design vs. Gamification

One last note: I am not advocating turning GenAI products into games, or gamifying them with badges and points. This would surely make the UX of LLMs much, much worse.

What I am proposing is playfulness across all aspects of these products, whatever you are using them for. As Jules says in Pulp Fiction: “Personality goes a long way.”

Playful design take advantage of humans natural curiosity and socialization instincts. It’s not about rewarding people for using your product, at least not in any traditional sense. It’s about making your product natural, pleasurable, and novel to use.

Sam Liberty is a gamification expert, applied game designer, and consultant. His clients include The World Bank, Click Therapeutics, and DARPA. He teaches game design at Northeastern University. He is the former Lead Game Designer at Sidekick Health.

--

--

Sam Liberty

Consultant -- Applied Game Design. Gamification expert. Clients include Click Therapeutics and The World Bank. Former Lead Game Designer at Sidekick Health.